DTrain

Distributed AI Model Training Platform — Powered by the Crowd

Final Year B.Tech Project — Department of Computer Science & Engineering

Hooghly Engineering & Technology College, Hooghly
Department of Computer Science & Engineering
Affiliated to Maulana Abul Kalam Azad University of Technology, West Bengal
2025–26

Demo

DTrain Demo Video

📽️ Replace the URL above with your actual video link (GitHub upload, Google Drive, or YouTube). GitHub supports direct .mp4 embeds in READMEs.

About DTrain

DTrain is a full-stack peer-to-peer distributed AI model training platform. It connects ML researchers and developers who need compute power with GPU owners who have idle machines — and handles everything in between: job analysis, pricing, payment escrow, execution, log streaming, and automatic payout.

Think of it as Airbnb for GPU compute. A user uploads their Python training code as a ZIP file, DTrain's two-layer AI pricing engine (rule-based scorer + Groq LLM) estimates job complexity and sets a flat price. The user reviews and pays. A worker's Electron desktop agent picks up the job, runs it inside an isolated Docker container, streams logs back live via Socket.IO, and uploads the trained model to Supabase when done. Payment settles automatically — the worker earns 80%, the platform takes 20%.

No cloud accounts. No infrastructure setup. No per-hour billing surprises.

The Problem It Solves

Pain Point	Reality Today
Cloud GPU costs	AWS p3.2xlarge costs ~₹280/hr. A BERT fine-tune can run 3–4 hours = ₹840–₹1120 for one job
Queue times	Colab Pro disconnects. Kaggle limits. University HPC has multi-day queues
Infrastructure complexity	Setting up CUDA drivers, Docker, cloud IAM, storage buckets takes hours
No live feedback	Cloud batch jobs give you nothing until they're done — or crashed
Wasted idle GPUs	Millions of consumer GPUs (gaming PCs) sit at 0% utilisation every night
No micropayment layer	Sharing compute ad-hoc has no billing, escrow, or trust mechanism

DTrain solves all six simultaneously.

Key Features

AI-powered job pricing — Groq LLM + rule-based static analyser reads your requirements.txt and training script to estimate complexity and assign a fair flat price from a ₹10–₹500 tier ladder. No surprises.
Draft → Publish flow — jobs are created as drafts with a price preview before any payment is taken. Users can review and cancel.
Stripe escrow — funds are reserved (not charged) on publish. Charge only happens on successful completion. Full refund on failure.
Real-time log streaming — every stdout/stderr line from the Docker container is pushed via Socket.IO to the user's browser as it happens.
Docker isolation — every training job runs in a fresh, sandboxed Docker container. No cross-job interference, no host machine access.
Electron desktop agent — a one-click Windows installer turns any gaming PC into a worker node. No terminal, no config files.
Automatic worker payout — when a job completes, the worker's in-platform wallet is credited instantly. Payout requests go through Stripe Connect.
Worker hardware detection — the Electron agent auto-detects OS, CPU, RAM, and GPU on registration.
80/20 revenue split — workers keep 80% of every job's tier price. Platform takes 20%.

System Architecture

The platform has four distinct application layers that communicate via REST APIs and WebSockets:

Component Responsibilities

Component	Tech	Purpose
`frontend/`	React 18, TypeScript, Tailwind, Vite	User dashboard — job submission, live log view, wallet top-up, model download
`backend/`	Express.js v5, Socket.IO, Mongoose	Central API — auth, job lifecycle, pricing, payment webhooks, log relay
`worker-ui/`	React, Tailwind, Vite	Browser UI for workers — view earnings, request payouts, set pricing preferences
`electron-worker/`	Electron, Node.js, Docker	Desktop agent — polls jobs, spins Docker containers, streams logs, uploads output

How It Works

User Flow (Step by Step)

Worker Flow (Step by Step)

Register — Worker opens the Electron app, which auto-detects device specs (OS, CPU, RAM, GPU via os module + Docker stats) and calls POST /api/worker/register.
Poll — Every ~5 seconds, the agent calls GET /api/worker/available-jobs?deviceId=.... This also acts as a heartbeat — if no poll arrives in 60s, the backend marks the worker offline.
Accept — Agent accepts the first available job. Backend atomically assigns it (prevents two workers grabbing the same job).

Execute — Agent downloads the ZIP from Supabase, extracts it to a temp folder, and runs:

docker run --rm -v /tmp/job_xyz:/workspace python:3.10-slim bash -c "pip install -r requirements.txt && python train.py"

Stream — Every line of stdout/stderr is POSTed to the backend, which relays it to the user via Socket.IO.
Complete — Output files are zipped and uploaded to Supabase. The agent calls POST /api/worker/complete-job. Backend settles payment and credits the worker wallet.

The AI Pricing Engine

Job pricing uses a two-layer scoring system to prevent both under-pricing heavy jobs and over-pricing trivial ones:

Layer 1 — Rule-based static analyser (paymentHelpers.js)

Scans requirements.txt (comment lines stripped) and the main Python file for known signals:

Signal	Score Added
`torch` / `pytorch` in requirements	+40 pts
`tensorflow` / `keras`	+40 pts
`transformers` / `diffusers` (Hugging Face)	+45 pts
`xgboost` / `lightgbm` / `catboost`	+20 pts
`scikit-learn`	+10 pts
`numpy` / `pandas`	+2 pts
`n_samples` ≥ 100,000 in code	+30 pts
`n_samples` ≥ 50,000	+20 pts
Epoch count ≥ 100	+15 pts
Multi-GPU / distributed training patterns	+20 pts

Layer 2 — Groq LLM reads the full script and returns its own complexity score 0–100.

The final score = max(rule_score, groq_score). This is then mapped to the tier ladder:

Score  0–10  → ₹10    Score 11–20  → ₹20    Score 21–30  → ₹30
Score 31–40  → ₹40    Score 41–50  → ₹50    Score 51–60  → ₹75
Score 61–70  → ₹100   Score 71–80  → ₹150   Score 81–88  → ₹200
Score 89–93  → ₹300   Score 94–97  → ₹400   Score 98–100 → ₹500

Revenue split:

Worker receives = tier_price × 0.80
Platform fee    = tier_price × 0.20

Payment & Wallet System

DTrain uses a reserve → charge model powered by Stripe to protect both parties:

User tops up wallet  →  Stripe payment intent
                              ↓
User publishes job   →  Funds RESERVED (not charged)
                              ↓
Worker accepts job   →  Job locked to worker
                              ↓
  ┌── Job SUCCESS ─────────────────────────────────────────┐
  │   Funds CHARGED · Worker wallet +80% · Platform +20%   │
  └────────────────────────────────────────────────────────┘
  ┌── Job FAILED / TIMEOUT ────────────────────────────────┐
  │   Reservation RELEASED · Full refund to user wallet    │
  └────────────────────────────────────────────────────────┘

Workers can request a payout from their in-platform wallet to their bank account via Stripe Connect at any time.

Performance & Cost Analysis

The following benchmarks compare estimated training time and cost across four compute setups for common ML workloads. All values are derived mathematically from published GPU TFLOPS specifications and real AWS spot pricing.

Methodology: Time = Total FLOPs ÷ (GPU TFLOPS × utilisation factor). GTX 1060 ≈ 4 TFLOPS @ 70% util · RTX 4090 ≈ 82.6 TFLOPS @ 85% util · AWS V100 ≈ 14 TFLOPS @ 90% util · DTrain assumes 3–5 mid-range workers (GTX 1080 Ti / RTX 3070, ~12 TFLOPS avg, 10% coordination overhead). AWS cost = time × ₹280/hr (p3.2xlarge spot). Local PC = electricity at ₹8/kWh. DTrain cost = flat tier price from the ₹10–₹500 platform ladder.

Training Time Comparison

Cost Per Job Comparison

DTrain Speed-up Factor vs Low-end PC

Full Numbers at a Glance

Model / Task	Low-end PC	✅ DTrain	High-end PC	AWS Cloud
Scikit-learn — time	1.4h	0.35h	0.25h	0.18h
Scikit-learn — cost	₹9	₹20	₹4	₹50
XGBoost — time	5.2h	1.1h	0.8h	0.55h
XGBoost — cost	₹33	₹40	₹12	₹154
PyTorch CNN — time	18.4h	4.1h	2.7h	1.8h
PyTorch CNN — cost	₹118	₹100	₹43	₹504
BERT fine-tune — time	38.0h	6.8h	5.2h	3.4h
BERT fine-tune — cost	₹243	₹200	₹83	₹952

Platform Feature Comparison

Platform	Cost (CNN job)	Setup Time	Live Logs	Auto Billing	Job Isolation
AWS EC2 p3.2xlarge	₹504	30–60 min	❌	❌	Full VM
Google Colab Pro	₹900/month flat	5–10 min	❌	❌	Shared runtime
Paperspace Gradient	₹400–₹800	15 min	Partial	❌	Container
Local machine only	~₹118 (electricity)	0 min	❌	N/A	None
✅ DTrain	₹100 flat	~2 min	✅ Real-time	✅ Stripe	Docker

DTrain's sweet spot is medium-to-heavy jobs (PyTorch CNNs, Transformers) where it is 4–6× faster than a low-end PC and 74–80% cheaper than AWS — with zero cloud account setup required.

Tech Stack

Backend (`backend/`)

Technology	Version	Role
	v20+	Runtime
	v5	REST API framework
	4.8	Real-time bidirectional communication (log streaming, worker status)
	Atlas + v9	Primary database — jobs, users, workers, billing, transactions
	v5 (ioredis)	Job queue, pub-sub messaging between backend and workers
	v1.2	LLM inference for job complexity analysis
	v22	Payment intents, webhooks, wallet management, Connect payouts
	v2	Object storage for ZIP uploads and model output files
	v9	Stateless auth — access tokens signed with `JWT_SECRET`
	v3	Password hashing
	v2	Multipart file upload handling
	v0.5	ZIP validation and file extraction on the server
	v1.10	HTTP request logging

Frontend (`frontend/`) — User Dashboard

Technology	Version	Role
	18.3	UI framework
	5.5	Type safety
	5.4	Build tool and dev server
	3.4	Utility-first styling
	4.8	Real-time log streaming and job status updates
	6.6 / 9.7	Embedded Stripe Checkout UI
	3.1	Job stats and earnings charts
	12	Page transition animations
	7.13	Client-side routing
	14	ZIP file drag-and-drop upload
	0.344	Icon set

Worker UI (`worker-ui/`) — Worker Dashboard

Technology	Version	Role
	18 / 5.5	UI framework
	5.4 / 3.4	Build + styling
	4.8	Live job status updates

Electron Worker Agent (`electron-worker/`)

Technology	Role
	Desktop app shell — system tray, native window, IPC
	Spawn Docker process, capture and stream stdout/stderr
	Detect CPU, RAM, platform for worker registration
	Run isolated training container for each job
	HTTP calls to backend API (job polling, log push, completion)
	Extract job ZIP before passing to Docker

Project Structure

 dtrain/
│
├──  backend                                       # Express.js REST API + Socket.IO server
│   ├──  middlewares
│   │   └──  authMiddleware.js                    # JWT verify → attaches req.user to every protected route
│   ├──  routes
│   │   ├──  JobRoutes.js                         # /create  /publish  /cancel  /list  /:id
│   │   ├──  PaymentRoutes.js                     # Stripe checkout, webhook handler, wallet top-up & payout
│   │   ├──  UserRoutes.js                        # /signup  /signin  /profile  /update  /delete
│   │   └──  WorkerRoutes.js                      # /register  /available-jobs  /accept  /push-log  /complete  /fail  /delete
│   ├──  schemas                                   # Mongoose data models (MongoDB)
│   │   ├──  BillingSchema.js                     # Per-job billing record — links user ↔ job ↔ worker
│   │   ├──  JobMetricsSchema.js                  # Training duration, resource usage snapshots
│   │   ├──  JobSchema.js                         # Core model — status enum, pricing, logs[], config{}
│   │   ├──  TransactionSchema.js                 # Wallet credit / debit ledger
│   │   ├──  UserSchema.js                        # email, passwordHash, walletBalance
│   │   └──  WorkerSchema.js                      # deviceId, gpu, currentStatus, walletBalance
│   ├──  utils
│   │   ├──  jwt.js                               # signToken / verifyToken helpers
│   │   ├──  paymentHelpers.js ⭐                  # Tier scoring engine + Groq LLM call + 80/20 split logic
│   │   ├──  redis.js                             # Redis publisher instance (job queue pub-sub)
│   │   ├──  stripeClient.js                      # Stripe SDK singleton
│   │   └──  supabaseClient.js                    # Supabase client (service role — ZIP upload & model fetch)
│   ├──  .gitignore
│   ├──  index.js ⭐                            # Entry point — Express app, Socket.IO setup, DB connect
│   ├──  package-lock.json
│   └──  package.json
│
├──  electron-worker                               # Desktop worker agent (Windows .exe)
│   ├──  assets
│   │   ├──  icon.ico
│   │   └──  icon.png
│   ├──  release                                   # Build output — do not edit manually
│   │   ├──  win-unpacked
│   │   │   ├──  locales                          # Electron i18n locale packs (auto-generated)
│   │   │   │   └──  *.pak  (55 languages)
│   │   │   ├──  resources
│   │   │   │   ├──  app.asar                     # Packaged Electron app bundle
│   │   │   │   └──  elevate.exe
│   │   │   ├──  DTrain Worker.exe               # Unpacked executable (for testing)
│   │   │   └──  *.dll / *.pak / *.bin           # Chromium runtime dependencies
│   │   ├──  DTrain Worker Setup 1.0.0.exe ⭐       # Distributable installer (electron-builder)
│   │   ├──  DTrain Worker Setup 1.0.0.exe.blockmap
│   │   ├──  builder-debug.yml
│   │   └──  builder-effective-config.yaml
│   ├──  .gitignore
│   ├──  main.js ⭐                              # Core agent — job polling, Docker spawn, log streaming, upload
│   ├──  package-lock.json
│   ├──  package.json
│   └──  preload.js                                 # Electron contextBridge — IPC between main ↔ renderer
│
├──  frontend                                      # User-facing React app  (port 5173)
│   ├──  public
│   │   ├──  Favicon.ico
│   │   ├──  logo.png
│   │   └──  logo1.png
│   ├──  src
│   │   ├──  components
│   │   │   ├──  ActiveWorkers.tsx               # Live map / list of online worker nodes
│   │   │   ├──  Dashboard.tsx ⭐                   # Job list, status filters, stats overview
│   │   │   ├──  Documentation.tsx               # In-app how-to guide
│   │   │   ├──  HeroSection.tsx                 # Landing page — hero, CTA, feature highlights
│   │   │   ├──  JobDetail.tsx ⭐                   # Live log stream, status timeline, model download
│   │   │   ├──  JobSubmission.tsx ⭐               # Multi-step: upload → AI pricing preview → publish & pay
│   │   │   ├──  PendingJobs.tsx                 # Queue view of jobs awaiting a worker
│   │   │   ├──  ProfileDropdown.tsx             # User menu — profile info, sign out
│   │   │   ├──  RunningJobs.tsx                 # Live view of currently processing jobs
│   │   │   ├──  SettingsModal.tsx ⭐              # Update name & email · permanent account deletion
│   │   │   ├──  SignIn.tsx                      # Login form with JWT storage
│   │   │   ├──  SignUp.tsx                      # Registration form
│   │   │   └──  Wallet.tsx ⭐                      # Balance card, Stripe top-up, transaction history
│   │   ├──  types
│   │   │   └──  index.ts                         # Shared TypeScript interfaces — Job, Worker, User, Billing...
│   │   ├──  App.tsx                             # Root component — router + auth guard
│   │   ├──  index.css                            # Tailwind base styles
│   │   ├──  main.tsx                             # React DOM entry point
│   │   └──  vite-env.d.ts                        # Vite env type declarations
│   ├──  .gitignore
│   ├──  eslint.config.js
│   ├──  index.html                           # HTML shell — Vite injects bundle here
│   ├──  package-lock.json
│   ├──  package.json
│   ├──  postcss.config.js
│   ├──  tailwind.config.js
│   ├──  tsconfig.app.json
│   ├──  tsconfig.json
│   ├──  tsconfig.node.json
│   └──  vite.config.ts
│
└──  worker-ui                                     # Worker browser dashboard  (port 3000)
    ├──  public
    │   ├──  Favicon.ico
    │   ├──  logo.png
    │   └──  logo1.png
    ├──  src
    │   ├──  components
    │   │   ├──  Documentation.tsx               # In-app guide for workers
    │   │   ├──  HeroSection.tsx                 # Worker landing / onboarding page
    │   │   ├──  JobDetail.tsx                   # Per-job detail with log preview
    │   │   ├──  PayoutRequest.tsx ⭐               # Payout form → POST /api/payment/payout-request
    │   │   ├──  PricingSettings.tsx             # Worker sets minimum accepted job price
    │   │   ├──  ProfileDropdown.tsx             # Worker menu — profile info, sign out
    │   │   ├──  RunningJobs.tsx                 # Active job monitor with live log preview
    │   │   ├──  SettingsModal.tsx ⭐              # Update name & email · delete user account + linked worker record
    │   │   ├──  SignIn.tsx                      # Worker login
    │   │   ├──  SignUp.tsx                      # Worker registration
    │   │   ├──  WalletCard.tsx                  # Earnings balance + payout trigger
    │   │   ├──  WorkerDashboard.tsx ⭐              # Earnings summary, recent jobs, wallet overview
    │   │   └──  WorkerRegistration.tsx ⭐          # First-time device registration — sends hardware specs
    │   ├──  types
    │   │   └──  index.ts                         # Shared TypeScript interfaces
    │   ├──  App.tsx                             # Router + auth guard
    │   ├──  index.css
    │   ├──  main.tsx
    │   └──  vite-env.d.ts
    ├──  .gitignore
    ├──  README.md
    ├──  eslint.config.js
    ├──  index.html
    ├──  package-lock.json
    ├──  package.json
    ├──  postcss.config.js
    ├──  tailwind.config.js
    ├──  tsconfig.app.json
    ├──  tsconfig.json
    ├──  tsconfig.node.json
    └──  vite.config.ts

Prerequisites

Ensure the following are installed and configured before running DTrain:

Runtime & Tools

Node.js v20 or later + npm v10+
Git
Docker Desktop — required on any machine running the Electron worker agent

Cloud Services (all have free tiers)

MongoDB Atlas — free M0 cluster is sufficient for development
Supabase — create a project and a storage bucket named jobs
Stripe — test mode keys work for the full flow
Groq — free API key for llama3-70b-8192 or compatible model
Redis — run locally via Docker (instructions below)

Environment Variables

Create a file named .env inside the backend/ directory. Never commit this file — it is already in .gitignore.

# ── Database ─────────────────────────────────────────────────────────────────
MONGO_URI=mongodb+srv://<username>:<password>@<cluster>.mongodb.net/?appName=DTrain

# ── Supabase (object storage for ZIPs and model outputs) ─────────────────────
SUPABASE_URL=https://<project-id>.supabase.co
SUPABASE_SERVICE_ROLE=eyJ...   # Service role key (NOT the anon key)

# ── Redis (job queue) ─────────────────────────────────────────────────────────
REDIS_URL=redis://127.0.0.1:6379

# ── Authentication ────────────────────────────────────────────────────────────
JWT_SECRET=<minimum-32-char-random-string>

# ── Stripe (payments) ─────────────────────────────────────────────────────────
STRIPE_SECRET_KEY=sk_test_<your-stripe-secret-key>
STRIPE_WEBHOOK_SECRET=whsec_<your-stripe-webhook-secret>

# ── Groq AI (job complexity pricing) ─────────────────────────────────────────
GROQ_API_KEY=gsk_<your-groq-api-key>

# ── Server ────────────────────────────────────────────────────────────────────
PORT=5000
FRONTEND_URL=http://localhost:5173

Where to find each value

Variable	Where to get it
`MONGO_URI`	MongoDB Atlas → Clusters → Connect → Drivers
`SUPABASE_URL`	Supabase dashboard → Project Settings → API
`SUPABASE_SERVICE_ROLE`	Supabase dashboard → Project Settings → API → `service_role` key
`JWT_SECRET`	Generate with: `node -e "console.log(require('crypto').randomBytes(32).toString('hex'))"`
`STRIPE_SECRET_KEY`	Stripe Dashboard → Developers → API keys
`STRIPE_WEBHOOK_SECRET`	Stripe Dashboard → Developers → Webhooks → signing secret
`GROQ_API_KEY`	console.groq.com → API Keys

Setup & Installation

Step 1 — Clone the repository

git clone https://github.com/<your-org>/dtrain.git
cd dtrain

Step 2 — Start Redis (via Docker)

DTrain uses Redis as a job queue and pub-sub channel. Start a Redis container:

docker run --name dtrain-redis -p 6379:6379 -d redis

Confirm it's running:

docker ps
# Should show dtrain-redis with status "Up"

To stop and remove later:

docker stop dtrain-redis && docker rm dtrain-redis

Step 3 — Configure environment variables

cd backend
cp .env.example .env   # if .env.example exists, otherwise create from scratch
# Fill in all values as described in the Environment Variables section above

Step 4 — Install all dependencies

Run from the project root:

# Backend
cd backend && npm install && cd ..

# User frontend
cd frontend && npm install && cd ..

# Worker browser UI
cd worker-ui && npm install && cd ..

# Electron desktop agent
cd electron-worker && npm install && cd ..

Step 5 — Supabase storage bucket

In your Supabase project, create a storage bucket named jobs and set it to public (so the signed URLs the backend generates are accessible to workers and users).

Running the Project

Open four separate terminal windows from the project root directory:

Terminal 1 — Backend API

cd backend
npm run dev

Expected output:

Server running on port 5000
MongoDB connected ✅
Redis connected ✅

Terminal 2 — Frontend (User Dashboard)

cd frontend
npm run dev

Open → http://localhost:5173

Terminal 3 — Worker UI

cd worker-ui
npm run dev

Open → http://localhost:3000

Terminal 4 — Electron Worker Agent

cd electron-worker
npm start

⚠️ Docker Desktop must be running before you launch the Electron agent. The agent spawns Docker containers to execute jobs. Without Docker running, job execution will fail immediately.

Running order matters

Docker Desktop  →  Redis container  →  Backend  →  Frontend / Worker UI  →  Electron Agent

API Reference

Authentication

Method	Endpoint	Auth	Description
POST	`/api/user/signup`	None	Register new user
POST	`/api/user/signin`	None	Login, returns JWT
GET	`/api/user/profile`	JWT	Get current user

Jobs

Method	Endpoint	Auth	Description
GET	`/api/jobs/`	JWT	List user's jobs (all statuses)
POST	`/api/jobs/create`	JWT	Upload ZIP, analyse, save as draft
POST	`/api/jobs/publish/:id`	JWT	Pay & publish draft to workers
GET	`/api/jobs/:id`	JWT	Get single job with logs
POST	`/api/jobs/cancel/:id`	JWT	Cancel a draft or pending job

Workers

Method	Endpoint	Auth	Description
POST	`/api/worker/register`	JWT	Register device + hardware specs
GET	`/api/worker/available-jobs`	None*	Poll for pending jobs (heartbeat)
POST	`/api/worker/accept-job`	None*	Claim a job
POST	`/api/worker/push-log`	None*	Stream a log line
POST	`/api/worker/complete-job`	None*	Mark job done + trigger payout
POST	`/api/worker/fail-job`	None*	Report failure + trigger refund

*Worker endpoints use deviceId for identification rather than JWT.

Payments

Method	Endpoint	Auth	Description
POST	`/api/payment/checkout`	JWT	Create Stripe payment intent
POST	`/api/payment/webhook`	Stripe sig	Handle Stripe events
GET	`/api/payment/wallet`	JWT	Get wallet balance + history
POST	`/api/payment/topup`	JWT	Add funds to wallet
POST	`/api/payment/payout-request`	JWT	Worker requests bank payout

Full Walkthrough

Manual walkthrough to test the full flow:

As a User:

Open http://localhost:5173 → Sign up
Go to Wallet → Top Up → use Stripe test card 4242 4242 4242 4242 (any future expiry, any CVC)
Go to Dashboard → Submit Job
Upload a ZIP containing your Python training script + requirements.txt
Enter your main filename (e.g. train.py)
Click Analyse — wait for the AI pricing engine to return a price
Review the price and click Publish — pay from your wallet
Navigate to the job detail page — wait for a worker to pick it up

As a Worker:

Open Docker Desktop — make sure it's running
Launch the Electron worker app (npm start in electron-worker/)
Open http://localhost:3000 → Sign up as a worker → Register device
The agent will automatically poll and pick up the published job
Watch logs stream in real time on both the worker agent and the user's job detail page
When done, the model is uploaded and your wallet is credited

Team

Role	Name	Roll No.	GitHub	LinkedIn
Team Member	Debjit Mitra	17600122046
Team Member	Soham De	17600122062
Team Member	Akash Poddar	17600122071
Team Member	Sourodip Ghosh	17600122040
Project Mentor	Ms. Agnimita Banerjee	Assistant Professor, Dept. of CSE	—	—

Roadmap

Potential improvements and features for future versions:

GPU-accelerated Docker containers (NVIDIA Container Toolkit)
Multi-worker job parallelism (data-parallel training)
Worker reputation and rating system
macOS and Linux Electron agent builds
Job template library (common training scripts)
Webhook notifications (email/Slack on job completion)
Admin dashboard for platform analytics
Worker GPU benchmarking before job assignment

License

This project was submitted as an academic final year B.Tech capstone project. All rights reserved by the authors and institution. Not licensed for commercial use or redistribution without explicit written permission from the team.

Built for reliability, distributed compute, and real-world ML accessibility

_{Made with ❤️ by Debjit, Soham, Akash & Sourodip · B.Tech Final Year Project 2025–26 · HETC, Hooghly}

Name		Name	Last commit message	Last commit date
Latest commit History 67 Commits
backend		backend
electron-worker		electron-worker
frontend		frontend
worker-ui		worker-ui
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

DTrain

Distributed AI Model Training Platform — Powered by the Crowd

Demo

Table of Contents

About DTrain

The Problem It Solves

Key Features

System Architecture

Component Responsibilities

How It Works

User Flow (Step by Step)

Worker Flow (Step by Step)

The AI Pricing Engine

Payment & Wallet System

Performance & Cost Analysis

Training Time Comparison

Cost Per Job Comparison

DTrain Speed-up Factor vs Low-end PC

Full Numbers at a Glance

Platform Feature Comparison

Tech Stack

Backend (backend/)

Frontend (frontend/) — User Dashboard

Worker UI (worker-ui/) — Worker Dashboard

Electron Worker Agent (electron-worker/)

Project Structure

Prerequisites

Environment Variables

Where to find each value

Setup & Installation

Step 1 — Clone the repository

Step 2 — Start Redis (via Docker)

Step 3 — Configure environment variables

Step 4 — Install all dependencies

Step 5 — Supabase storage bucket

Running the Project

Terminal 1 — Backend API

Terminal 2 — Frontend (User Dashboard)

Terminal 3 — Worker UI

Terminal 4 — Electron Worker Agent

Running order matters

API Reference

Authentication

Jobs

Workers

Payments

Full Walkthrough

Team

Roadmap

License

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Backend (`backend/`)

Frontend (`frontend/`) — User Dashboard

Worker UI (`worker-ui/`) — Worker Dashboard

Electron Worker Agent (`electron-worker/`)

Packages